Research

At the AGML Research Center, our mission is to advance the frontiers of artificial intelligence through a harmonious integration of algorithmic foundations, graph-based methodologies, and deep learning innovations. Our research spans four interconnected pillars: Algorithms and Graphs, where we design efficient computational geometry solutions and data structures; Graph Neural Networks, which model complex relational data; Knowledge Graphs, which structure and enrich semantic information; and Deep Learning, which powers predictive models with a strong emphasis on interpretability and explainability. By bridging rigorous theoretical computer science with practical, deployable AI systems, we tackle some of the most challenging problems in machine learning—from scalable graph processing and spatial data analysis to trustworthy AI in high-stakes domains like healthcare and finance. Our interdisciplinary approach ensures that every advancement is not only scientifically robust but also directly applicable to real-world challenges. Beginning with fundamental concepts, we build upon them to develop sophisticated models that analyze complex problems, ultimately leading to the discovery of exact or approximate solutions and predictions. Through this synergy of algorithms, graphs, and learning, we strive to unlock the full potential of AI and contribute meaningfully to the broader scientific community and society.

At the heart of every intelligent system lies a robust algorithmic backbone. Our research in this pillar is deeply rooted in Computational Geometry—the study of algorithms for solving geometric problems concerning points, lines, polygons, and higher-dimensional structures. We design and optimize provably efficient algorithms for spatial data processing, which are critical for applications such as autonomous navigation, 3D computer vision, spatial databases, and physical simulations.

Beyond geometry, our expertise extends to advanced Data Structures that are fundamental for managing and querying complex, high-dimensional datasets. From spatial partitioning trees (like KD-trees and R-trees) to dynamic graph indices and probabilistic data structures, we engineer the underlying frameworks that make modern machine learning pipelines computationally tractable. We explore how tailored data structures can dramatically accelerate nearest neighbor searches, subgraph matching, and hierarchical clustering—operations that are bottleneck steps in many AI systems.

Furthermore, we bridge the gap between theoretical computer science and applied machine learning by developing Approximation and Heuristic Algorithms. These techniques are essential for tackling NP-hard problems that frequently arise in large-scale graph processing and machine learning, such as community detection, influence maximization, and feature selection. While approximation algorithms provide guaranteed performance bounds, heuristic approaches offer practical, scalable solutions that work exceptionally well in real-world scenarios where optimal solutions are computationally infeasible. By combining these algorithmic strategies with geometric insights and cutting-edge data structures, we ensure that our machine learning models are not only accurate but also computationally feasible, scalable, and robust for real-world, high-dimensional data. Ultimately, our work in algorithms and graphs provides the essential mathematical and computational foundation upon which all other AI research at AGML is built.

Imagine information as a web of interconnected nodes and edges. Graph Neural Networks are a powerful type of deep learning model that excels at analyzing such complex relationships. They work by iteratively sharing meaning between nodes, allowing them to understand the context and connections within the data. This makes GNNs invaluable for tasks like sentiment analysis, bioinformatics, and even remote sensing. Graph Machine Learning provides a new set of tools for processing network data and leveraging the power of the relation between entities that can be used for predictive, modeling, and analytics tasks. Most of relational structures in real world do not satisfy the simplicity required to be analyzed by neural networks. When neural networks fail, graph learning prevails!

GNNs have also made significant inroads in Natural Language Processing (NLP). Words can be represented as nodes, and their relationships (e.g., grammatical dependencies, semantic similarities) as edges, specifically, with the Relational GCN model. GNNs can then analyze these connections to perform tasks like sentiment analysis, machine translation, and question answering. This opens up exciting possibilities for understanding and interacting with language in new ways.

Temporal Graph Machine Learning is a specialized field that analyzes dynamic graphs, where data changes over time. Temporal Graph ML allows us to predict future properties of nodes and links within these evolving networks, finding applications in areas like social networks and financial markets.

In the realm of knowledge representation, knowledge graphs have emerged as a powerful paradigm. They leverage graph-structured databases to encode entities (think objects, events, or ideas) and the relationships that bind them. Unlike traditional databases, knowledge graphs don’t merely store information; they capture the rich semantics that connect data points. This enables advanced tasks like reasoning and inference, allowing machines to glean implicit knowledge from the explicit relationships in the graph. Imagine a knowledge graph about a historical figure; it wouldn’t just list their birthdate, but connect them to their contemporaries, their influence on specific events, and the broader historical context. This structured approach empowers machines to move beyond simple data retrieval and delve into the intricacies of interconnected knowledge.

Knowledge graphs (KGs) and large language models (LLMs) are becoming a powerful duo in the world of AI. KGs act as structured repositories of information, like meticulously organized libraries. LLMs, on the other hand, excel at understanding and generating natural language. This complementary nature creates exciting possibilities. LLMs can leverage KGs to ground their understanding in factual knowledge, making their responses more accurate and relevant. Conversely, LLMs can analyze massive amounts of text to populate and enrich KGs, filling in missing information and uncovering new relationships. This KG-LLM teamwork fuels applications across various domains. Imagine a recommendation system that not only considers your past purchases but also taps into a knowledge graph to understand product features, user reviews, and complementary items. This synergy between structured knowledge and natural language processing unlocks a future of more intelligent and informative AI systems.

Deep learning forms the backbone of many of our research endeavors. This powerful technique uses artificial neural networks to learn from data and make predictions. We leverage deep learning in various ways, including: Large Language Models (LLMs) in Healthcare Assistant Development. LLMs are a type of deep learning model trained on massive amounts of text data. We utilize LLMs to analyze medical information and support tasks like diagnosis, treatment, and prognosis within healthcare assistant systems.

As AI becomes increasingly integrated into our lives, it’s crucial to understand how these models make decisions. Explainable Artificial Intelligence (XAI) and Interpretable Machine Learning (IML) focus on developing techniques and tools that provide transparency and accountability for AI systems. Our research in this area aims to design novel algorithms and methods that can “explain” the behavior of complex AI models, allowing humans to trust, validate, and improve their performance. This is particularly important in high-stakes applications like healthcare and finance. Counterfactual Explanations for Deep Learning Models: While deep learning models are powerful, they can sometimes be difficult to understand. Our research in this area focuses on developing methods to create “counterfactual explanations,” allowing us to understand the causal relationships learned by these models and improve their interpretability.